A Comparative Study of Machine Learning and Deep Learning Models for Automatic Parkinson's Disease Detection from...

Background: Parkinson's disease (PD) is one of the most prevalent, widespread, and intricate neurodegenerative disorders. According to the experts, at least 1% of people over the age of 60 are affected worldwide. In the present time, the early detection of PD remains difficult due to the absence of a clear consensus on its brain characterization. Therefore, there is an urgent need for a more reliable and efficient technique for early detection of PD. Using the potential of electroencephalogram (EEG) signals, this study introduces an innovative method for the detection or classification of PD patients through machine learning, as well as a more accurate deep learning approach. Methods: We propose an innovative EEG-based PD detection approach by integrating advanced spectral feature engineering with machine learning and deep learning models. Using (a) the UC San Diego Resting State EEG dataset and (b) IOWA dataset, we extract a standardized EEG feature from five key frequency bands-alpha, beta, theta, gamma, delta (α,β,θ,γ,δ) and employ an SVM (Support Vector Machine) classifier as a baseline, achieving a notable accuracy. Furthermore, we implement a deep learning classifier (CNN) with a complex multi-dimensional feature set by combining power values from all frequency bands, which gives superior performance in distinguishing PD patients (both with medication and without medication states) from healthy patients. Results: With the five-fold cross-validation on these two datasets, our approaches successfully achieve promising results in a subject dependent scenario. The SVM classifier achieves competitive accuracies of 82% and 94% in the UC San Diego Resting State EEG dataset (using gamma band) and IOWA dataset, respectively in distinguishing PD patients from non-PD patients in subject. With the CNN classifier, our model is able to capture major cross-frequency dependencies of EEG; therefore, the classification accuracies reach beyond 96% and 99% with those two datasets, respectively. We also perform our experiments in a subject independent environment, where the SVM generates 68.09% accuracy. Conclusions: Our findings, coupled with advanced feature extraction and deep learning, have the potential to provide a non-invasive, efficient, and reliable approach for diagnosing PD, with further work aimed at enhancing feature sets, inclusion of a large number of subjects, and improving model generalizability across more diverse environments.